Google: AI-powered malware uses large language models to self-modify

Google's Threat Intelligence Group said Wednesday that cybercriminals and state-backed actors are increasingly embedding generative artificial intelligence into active malware, creating self-modifying programs that use large language models during execution to rewrite code, evade detection, and tailor attacks in real time. The report identifies recent families and samples observed this year, including PROMPTFLUX, PROMPTSTEAL, PROMPTLOCK, FruitShell, and QuietVault, and warns the technique marks a shift from AI as a productivity aid to AI as an operational capability in attacks.

GTIG investigators described a "just-in-time" model in which malware calls external LLMs mid-execution to generate or obfuscate code on demand. PROMPTFLUX, for example, has used the Gemini API to periodically rewrite its VBScript source and save obfuscated variants to achieve persistence and evade signature-based antivirus. PROMPTSTEAL has queried open-source models via Hugging Face to generate one-line Windows commands that harvest system files and documents. Other observed tools include FruitShell, a PowerShell reverse shell that enables remote command execution, and QuietVault, a JavaScript credential...